CSSE463: Image Recognition CSSE463: Image Recognition Day 25Day 25 Today: introduction to object recognition: Today: introduction to object recognition:

template matchingtemplate matching

Template matching: a simple method for object Template matching: a simple method for object detectiondetection

Questions?Questions?

Template matching (Sonka, 6.4)Template matching (Sonka, 6.4) Idea: you are looking for an exact match of Idea: you are looking for an exact match of

an object (described by a sub-image, a an object (described by a sub-image, a templatetemplate) in an image) in an image

Ideal world: it matches exactlyIdeal world: it matches exactly

Template matching (Sonka, 6.4)Template matching (Sonka, 6.4) Algorithm:Algorithm:

Evaluate a match Evaluate a match criterion at every image criterion at every image location (and size, location (and size, reflection, and rotation, if reflection, and rotation, if those variations are those variations are expected)expected)

A “match” is a local A “match” is a local maximum of the criterion maximum of the criterion above a thresholdabove a threshold

Q1

Template matching (Sonka, 6.4)Template matching (Sonka, 6.4)One match criterion:One match criterion:

Correlation between the template and the Correlation between the template and the image.image.

We are just using the template as a filter!We are just using the template as a filter!Simplistic implementationSimplistic implementationSmarter implementationSmarter implementation

CorrelationCorrelation Just the dot product between Just the dot product between

the template and a the template and a neighborhood in the image. neighborhood in the image.

Idea: high correlation when Idea: high correlation when the template matches.the template matches.

Demo Demo

CorrelationCorrelation Just the dot product between Just the dot product between

the template and a the template and a neighborhood in the image. neighborhood in the image.

Idea: high correlation when Idea: high correlation when the template matches. the template matches.

Problem: Problem: alwaysalways high high correlation when matching correlation when matching with a plain bright regionwith a plain bright region

Q2-3

CorrelationCorrelation Just the dot product between Just the dot product between

the template and a the template and a neighborhood in the image. neighborhood in the image.

Idea: high correlation when Idea: high correlation when the template matches. the template matches.

Problem: Problem: alwaysalways high high correlation when matching correlation when matching with a plain bright regionwith a plain bright region

Solution: Normalize the Solution: Normalize the template and each region by template and each region by subtracting each’s mean from subtracting each’s mean from itself itself before before taking dot producttaking dot product

Q4-5

Other matching algorithmsOther matching algorithmsChamfering (Hausdorff distance):Chamfering (Hausdorff distance):

http://www.cs.cornell.edu/~dph/hausdorff/hausdorff1.htmlhttp://www.cs.cornell.edu/~dph/hausdorff/hausdorff1.html

Springs and templates (Crandall and Springs and templates (Crandall and Huttenlocher)Huttenlocher) http://www.cs.cornell.edu/~dph/papers/cvpr07.pdfhttp://www.cs.cornell.edu/~dph/papers/cvpr07.pdf

Watershed segmentation (Sonka 6.3.4)Watershed segmentation (Sonka 6.3.4)